Management’s Best Estimate of Loss Reserves Abstract

advertisement
Management’s Best Estimate of Loss Reserves
Rodney Kreps
Guy Carpenter Instrat
Abstract
An economically rational way for management to set reserve estimates is to utilize the
future change in the value of the company as a statistical decision function and then to
choose the reserve estimate so as to minimize the average value of this function. The
mean of the reserve distribution is almost surely too low as an outcome.
Introduction
Management is required1 to provide a best estimate of loss reserves. In the opinion of
this author, actuarial practices2 strongly suggest the estimate be the mean value of the
distribution of loss reserves. It will be argued that the problem of the estimate is best
approached by statistical decision theory, and that all the usual statistical estimates can be
produced in such a fashion. Further, there is an economic basis for choosing a decision
function, which then determines the estimate. Desirable characteristics for a decision
function are discussed, and a candidate function is proposed. A simplified example and a
spreadsheet are provided. One general conclusion that emerges is that the mean is
probably not a good estimate, as it is almost surely low.
Statement of Statutory Accounting Principles
The SSAP #55 effective January 1, 2001 says3 in part “For each line of business and for
all lines of business in the aggregate management shall record its best estimate of its
liabilities …” and “Management’s analysis of the reasonableness of … reserve estimates
shall include an analysis of the amount of variability in the estimate.” Not to put too
much into a single word, but please note that it is “in the estimate” rather than “of the
estimate.” The author believes that actuaries have tended to place too much attention on
the differences between different estimates and not enough on the variability of actual
results.
SSAP 55 goes on to say “Management’s range [of estimates] shall be realistic and,
therefore, shall not include the set of all possible outcomes but only those outcomes that
1
Statutory Statement of Accounting Principles #55 effective January 1, 2001
ASOP # 36, especially section 3.6.3
3
This and subsequent quotes in this section are taken from SSAP 55, page 55-6, sections 10 and 11.
2
are considered reasonable.” In other words, weight scenarios by their probabilities: use a
distribution. But how?
In the next section SSAP 55 says that in the case of a range with all values equally likely,
choose the middle of the range; but if the equally likely values do not form a range
“management should determine its best estimate of the liability.” Again, there is not a lot
of help here on how to actually do it. Let us see what the actuaries have to say.
Actuarial Standard of Practice
ASOP 36 – Statements of Actuarial Opinion Regarding Property/Casualty Loss and Loss
Adjustment Expense Reserves - has quite a bit to say about reserves and uncertainty. We
will assume that the hard work of evaluating trends, court climates, and other sources of
both process and parameter uncertainty has been done and that we have a best candidate
loss reserve distribution which includes all the outcomes and best estimates of their
probabilities.
This is, of course, a major assumption seldom made explicit in practice but often used
implicitly. For example, section 3.6.3 is entitled “Expected Value Estimate” and says “In
evaluating the reasonableness of reserves, the actuary should consider one or more
expected value estimates of the reserves, except when such estimates cannot be made
based on available data and reasonable assumptions.” Expected value, apart from the fact
that you never actually expect to see it happen, is the mean of a distribution. So “one or
more expected value estimates” is saying “find different ways of getting the mean of our
unknown but implicitly present distribution.”
The same section goes on to say “Other statistical values such as the mode (most likely
value) or the median (50th percentile) may not be appropriate measures for evaluating loss
and loss adjustment expense reserves, such as when the expected value estimates can be
significantly greater than these other estimates.” Here the author sees the innate,
carefully cultivated, and experientally substantiated conservatism of the actuaries
expressed as “Let’s go for the higher value.” Curiously enough, this paper will argue that
the mean itself is too low.
The section continues “The actuary may use various methods or assumptions to arrive at
expected value estimates. In arriving at such expected value estimates, it is not necessary
to estimate or determine the range of all possible values, nor the probabilities associated
with any particular values.” So, although ASOP wants mean value estimates, it does not
want a distribution. Most of the techniques for doing reserve estimates, even including
those which are used for sparse or missing data, have an underlying statistical model and
an implied distribution. We are invited to use the models and forget the distributions.
The next section 3.6.4 is entitled “Range of Reasonable Reserve Estimates.” It begins
“The actuary may determine a range of reasonable reserve estimates that reflects the
uncertainties associated with analyzing the reserves. A range of reasonable estimates is a
range of estimates that could be produced by appropriate actuarial methods or alternative
sets of assumptions that the actuary judges to be reasonable.” Clearly, something like
this needs to be in the ASOP in order to give the actuary room not to be forced by a
formula. This is saying, look at least implicitly at several reasonable distributions
(alternatively, models), get the mean from each one, and make a weighted choice. The
author would prefer that the judgment calls be in the creation of the best predictive
distribution. For example, if there are two equally valid distributions, weight them
equally. The mean will be the average of the individual means. Of course, this does
require going from no distributions to three.
This same section continues “The actuary may include risk margins in a range of
reasonable estimates, but is not required to do so, except as required by ASOP No. 20. A
range of reasonable estimates, however, usually does not represent the range of all
possible outcomes.” So, while allowing that there really is a distribution with outcomes
of differing probabilities, the ASOP wants to make sure that the reserve estimate is well
inside the range while not giving any guidance on how to get to a risk margin or what
might be appropriate. Presumably, the margin should be for the risk that there will be, in
the words of section 3.3.3, an “amount of adverse deviation that the actuary judges to be
material with respect to the statement of actuarial opinion..” Unless the distribution is
very narrow, this seems quite likely to be the case.
Section 3.6.5 on “Adverse Deviation” makes the same point, but does not suggest a risk
margin. It only says “The actuary should consider whether the future paid amounts are
subject to significant risks and uncertainties that could result in a material adverse
deviation.” Section 4.6(g) says that it is up to the actuary to include mention of this in the
report . In addition, section 4.8 says “An actuary must be prepared to justify the use of
any procedures that depart materially from those set forth in this standard and must
include, in any actuarial communication disclosing the results of the procedures, an
appropriate statement with respect to the nature, rationale, and effect of such departures.”
Perhaps this justification could be “and I plunked down another 35% because this line is
all over the place.”
The author’s sense of the ASOP writing on uncertainty is that it gives management some
idea of how much wiggle room there is in the creation of management’s best estimate, at
least according to the actuaries. But none of this actually gives an economic basis for
how the estimate should be made.
Statistical Decision Theory – the short version
An economic basis for the creation of an estimate needs a way to combine a probability
distribution of outcomes with an economic function describing the pain that will be felt
when the realized random outcome differs from the estimate. The simplest recipe is to go
for Least Pain, as follows:
(1) Create a pain function based on economic reality. This function will be the economic
decision function which will drive the results. “Pain function” actually is the technical
term because this function is meant to represent how unpleasant adverse outcomes of
various sorts may be. This function will depend on the estimate and the random variable
representing possible outcomes. Typically it will be zero when random outcome and
estimate are equal.
(2) Average the pain function over the probability distribution of outcomes for every
fixed estimate.
(3) Find the value of your estimate which makes the average pain smallest.
It is easy enough to see that such a prescription satisfies SSAP 55, and that the
description of the pain function represents management’s logic in creating the estimate.
The context is in hand is setting the reserves and then a year later making adjustments for
development on old years. More precisely, we assume that the assessment a year later is
“correct” and represents a random realization of the underlying distribution.
What we will show next is that all the usual estimates can be represented as being derived
from pain functions. The comparison of pain functions gives us a way to speak about the
relevance of the estimates in a business and economic context. Pain functions can be
thought of as negative utility functions.
Following that we will argue for the general characteristics of a pain function for loss
reserving.
Mathematical Representation of the recipe
The general case is that we have a probability density function f(x) with support from 0 to
infinity. We also have a pain function p(m,x) which is a function of the estimate m and
the random variable x. We denote the average of p over f as P(m):

P  m   E  p    p  m, x  f  x  dx
0
It is reasonable to ask that the integral exists, and that p >= 0 everywhere. We want m to
be such that P(m) is the smallest, so we choose the value for m which makes

d

0
P  m  
p  m, x  f  x  dx
dm

m
0
Sometimes in practice p will be discontinuous at x = m. In that case define
for x  m
 p (m ,x)
p(m ,x) =  for x  m
 p+ (m ,x)
which makes
m

dP(m)
 p- (m ,x)
 p (m ,x)
=
f(x)dx   +
f(x)dx
dm
m
m
0
m
 p  m, m  p  m, m
Usually the last difference is zero and in fact the individual terms are usually zero.
Notice that the scale and absolute value of P(m) do not enter into the calculation for m.
You can add a constant and multiply by any constant and m does not change. The pain
functions given are the only ones known to the author which give the usual statistics for
all distributions.
Example: the mean
For the mean, the pain function is a quadratic about the estimate:
2
p  m, x    m  x 
The average pain is

P  x     m  x  f  x  dx
2
0
In this particular case, we can do the integrals in terms of the mean and variance of the
distribution:

P  m     m 2  2mx  x 2  f  x  dx
0
 m 2  2m * mean  Var  mean 2 
 Var   m  mean 
As a function of m, this clearly has a minimum when m equals the mean of the
distribution.
2
At this point, we should pause and ask ourselves “Why do I want a quadratic decision
function? What is so good about squared dollars?” The symmetry of the pain function
about the estimate implies that for the reserves to come in lower than our estimate is as
bad as having them come in higher. The quadratic form implies that two dollars low is
four times as bad as one dollar low.
Example: the median
For the median, the pain function is linear about the estimate with equal slope on both
sides, and with a discontinuity in the derivative at x = m:
p  m, x   abs  m  x 
for x  m
m  x

for x  m
x  m
Although this function is still symmetric about the estimate, it says that it is dollars,
rather than squared dollars, that are of interest; and that two dollars off is only twice as
bad as one dollar off. This has some plausibility.
For the evaluation, the partial derivative is
 p(m ,x)  1
=
m
-1
for x < m
for x > m
So the equation to be solved for m is
m

0=

f(x)dx   f(x)dx
0
m
 F  m   1  F  m  
where F(x) is the cumulative distribution function. This requires
F m  1
2
That is to say, m is the median.
Example: the fixed percentile
For an arbitrary fixed percentile, the pain function is linear about the estimate, but with
different slopes on the high and low side:
for x  m
m  x
p  m, x   
for x  m
  x  m 
Again it is dollars that are of interest but here it is a factor of  worse for x to be high
rather than low. Take  to be some constant, say 3.
For the evaluation, the partial derivative is
 p(m ,x)  1
=
m
-
So the equation to be solved for m is
m
0=
for x < m
for x > m


f(x)dx    f(x)dx
0
m
 F  m    1  F  m  
where again F(x) is the cumulative distribution function. This requires

F m 
 1
That is to say, m is the   +1 percentile value. For  = 3, this is the 75th percentile.
It will be argued that a decision function that gives more weight to the high side than the
low is desirable for loss reserving. It is usually worse to come in above your estimate
than below it.
Example: the mode
Here the decision function is one outside of some (preferably small) interval around the
estimate, and zero within it. The economic interpretation of this decision function is that
for the reserves to come in outside of this interval is equally bad no matter where it
happens. This means that high or low, just outside or very far away are all the same.
This does not seem reasonable. On the other hand, this will provide the single best guess
for an interval of given size to contain the result.
The pain function is
for x  m  
1

p(m ,x) = 0
for m    x  m  
1
for x  m  

and 2 is the size of the interval. The average pain is the probability that the random
variable is realized outside of the interval:
P  m   F  m     1  F  m    
Setting the derivative to zero,
df  m 
dm
For small interval, this says that the density function is a maximum at m. That is, m is
the mode.
0  f  m     f  m     2
Fundamental considerations
It is clear that all of the usual estimates can be phrased as resulting from a particular
choice of decision function and that infinitely many decision functions are possible.
What is needed is to construct the decision function from the economic or other forces
which impact the entity setting the reserve levels. This will be the appropriate decision
function.
One possibility is a purely subjective estimate of how, say, the CFO feels about various
sizes of future difference of reserves from the stated estimate. Slightly better would be
to use the reserves committee as input, in a Delphi method.
Another possibility is to examine the fundamental economic consequences which result
from the reserves (at least as appearing in next year’s Annual Statement) coming in
different from the reserve level which is currently set. A good candidate for the decision
function is the decrease in the net economic worth of the company as a result of the
reserve changes. While estimates of this may involve subjective judgments, at least
something of definite and measurable economic value is being considered.
Interested parties who may affect the economic value of the company include
policyholders, stockholders, agents, regulators, rating agencies, IRS, investment analysts,
and lending institutions.
If the reserves come in slightly higher than the estimate, there perhaps is not much market
reaction. The industry as a whole has had the reputation of being under-reserved. It may
also be that some managements will like being slightly under-reserved because then they
are able to have overstated earnings the previous year. However, if the increase in
reserve levels is significant enough that surplus is significantly impacted then a number
of effects come into play: The capacity to write business is impaired; the firm’s credit
rating may become impaired, increasing the cost of capital; rating agencies, investor
analysts, and the IRS become more concerned; future renewals (the “goodwill” of the
firm) become more problematical. And if the change is large enough then IRIS tests
begin to be triggered and regulatory authorities are involved, which definitely will
decrease the value of the firm. The economic consequences would seem to be rapid and
non-linear in the reserve increase.
On the other hand suppose the reserves come in significantly below the estimate. This
means that the company has been over-reserved and consequently4 is less competitive
than it could have been; that the IRS has ammunition for its audit; that dividends could
have been larger; participatory plans could have been more generous; that there is a
danger of losing future business from over-pricing. These effects would seem to be less
immediately significant than the results from under-reserving, at least in the short term.
Each of these situations will generate a negative effect on the net worth of the company
compared to its value if the reserves came in as stated. However, intuition suggests that
the effect will be much stronger on the under-reserving side than on the over-reserving
side, and will be non-linear, especially as particular analyst, rating agency, and regulatory
tests reach trigger points.
The immediate consequence is that a symmetric pain function such as that for the mean
would be over-estimating the negative impact of reserve decreases and consequently the
estimate is intrinsically too low.
An approximation to the correct function
As a crude approximation which has some of the properties just suggested, consider a
decision function which is quadratic around the estimator but linear (using the tangent to
the parabola) at some value below it. Call it “semi-quadratic.” This function is
mathematically well-behaved, as the value and the first derivative are both continuous.
This function will also satisfy that being high (in the outcome) is never better than being
low and that the high side is quadratic while the low side is linear. The choice of distance
below determines the slope of the line; the closer it is to the estimator the smaller the
slope.
We make the decision function have the dimensions of dollars to mimic the economic
value. S is the company surplus, since that is the appropriate scale for many tests.
The decision function is
4
2  m-x    S

p(m ,x) =   x-m  2

 S
for x  m   S
for x  m   S
Assuming that the pricing and reserving actuaries actually talk with each other
The parameter  is dimensionless and reflects management attitude. The pain is equal to
S when the outcome is S above or below the estimator. A small  reflects a relatively
low pain for over-reserving and conversely a higher pain for under-reserving.
An  of 3% is used below, which says that a reserve change of 3% of surplus downward
is the same pain as the same change upward. However, 10% under-reserved is about
twice as bad as 10% over-reserved; 20% under-reserved is about 3.6 times as bad as 20%
over-reserved. Again, as  gets smaller management is less tolerant of under-reserving.
 -2
 p(m ,x) 
=  2  x  m
m
 S

The partial derivative is
for x  m  S
for x  m  S
and the equation to be solved is
m  S
0=-

0

f(x)dx 
1
 x-m  f(x)dx
 S m S


1
F m   S  
Or,
 mean  F1  m   S    m 1  F  m   S  
S 
Where F1(x) is the integral of xf(x) – the first moment distribution.
In order to work entirely with dimensionless variables, it is convenient to measure the
estimate and the mean in units of surplus. Then the above equation holds with S = 1.
Just to make explicit the kind of results that might be seen, as an example take F(x) to be
a lognormal distribution with known mean and coefficient of variation c. Both F(x) and
F1(x) can be explicitly calculated in terms of the normal distribution function
x
z2

1
N  x 
 e 2 dz :
2 
We have
and
with the usual
 ln  x    
F  x  N 




 ln  x   

F1  x   mean N 
  .



  ln1  c2  and
  ln  mean    2 .
2
For a company which has a mean reserve to surplus ratio of 3.5 with a coefficient of
variation in the reserves of 10%, an  of 3% results in setting the reserve estimator at
11.5% above the mean. This estimator is at about the 87.2% level of the reserve CDF.
See the accompanying worksheet SemiQuadraticExample.xls for details and other values
of .
Why bother?
Almost everything actuarial that goes into actually setting reserves is assumed in this
discussion. In particular, the explicit existence of a distribution is problematical. Not
impossible, just hard.
In many companies the current reserve-setting process probably already done on a leastpain basis. However, the pain is not future pain but present pain. Reserves are set with
an eye on what was set in the past and on current analyst expectations. The reserve
committee is always playing catch-up. The procedure discussed here assumes that the
reserves are set on old years (how the random variable comes in) with no concern for
current politics – clearly a naïve assumption.
Also, the author is not aware of a line of business decision function that would make
sense. Perhaps some of the capital allocation methodologies could be helpful.
Similarly, no one actually knows how the value of the company will decrease; but
experienced players have some sense of it. There have been enough examples in the last
few years to show that reserve changes can have significant impact. It is also clear that
what may impact a given company’s pain function may be quite different from another’s,
and that the emphasis may very well change from year to year. Allowing this would be a
problem for regulators.
Still, it would be a useful exercise for the reserve committee at a company to get together
and try to build, even crudely, their pain function for the year. They could perhaps begin
with some standardized event (lose 10% of surplus) which has enough pain to work as a
comparison with other possibilities, and then fill out both the high and low sides at a
convenient and realistic set of values5. Then they could ask the actuary to do the
numerical calculation for the estimate, and have a much better idea of what the increase
in average pain would be from using an estimator not at the minimum if they choose to
do so. And, in the process they would come to be able to explain how they arrived at
their estimate. Another interesting exercise6 would be to put the management incentive
plan as an input to the pain function.
All of this makes reserving by formula (e.g., use the mean) impossible. But it really is,
anyway.
5
For example, let loss 10% of surplus = 1 pain unit then make up the rest: lose 25% = 10; lose 20% = 8;
lose 15% = 3; lose 5% = 0.5; lose 0 = 0; gain 5% = 0; gain 10% = 0.5; gain 15% = 1; gain 20% = 2.
6
This is probably implicitly happening already.
Download